A Novel Method for Enforcing Exactly Dirichlet, Neumann and Robin Conditions on Curved Domain Boundaries for Physics Informed Machine Learning

The decomposition of compatibility constraints at vertices is careful and complete. When two Neumann boundaries meet at a vertex, the authors derive explicit constraints on the cross-derivatives $V_{\xi\eta}(1,1) = V_{\eta\xi}(1,1)$ that must hold for the PDE solution, using the coupling coefficients $S_{BC}$ and $S_{CD}$. The numerical verification is convincing: boundary condition errors achieve machine precision ($10^{-15}$-$10^{-16}$) consistently across linear Helmholtz, nonlinear Helmholtz, and moving-boundary heat equation problems on multiple non-trivial geometries including crescent and star-shaped domains. The TFC formulation correctly eliminates the free function contributions on boundaries through the projector $P$.

The scope is limited to quadrilateral topologies, which the authors acknowledge but nonetheless restricts applicability to problems amenable to this parameterization. The complexity of implementing the 4-step procedure (identifying constrained variables, constructing transfinite interpolants, preliminary TFC form, then updating unknown terms) grows substantially with mixed BC types—the Robin-Robin intersection case requires tracking 9 distinct constraint types (Section 2.4.2). The method requires explicit parametric descriptions of all boundary curves, which may not be available for implicitly defined geometries. The cancellation error issue noted in Remark 2.10, where $(g - Pg)$ can suffer from numerical cancellation at isolated boundary points, suggests the 'exact' enforcement is theoretically rigorous but numerically delicate; the authors partially address this with auxiliary collocation constraints $g=0$ on Dirichlet boundaries.

The evidence strongly supports the claim of machine-accuracy boundary enforcement across diverse geometries and PDE types. The comparison to 'soft' enforcement methods is discussed in the literature review (Section 1), citing gradient-flow pathologies and weight sensitivity issues in standard PINNs. However, no direct numerical comparison between the proposed exact enforcement and penalty-based methods is presented in the results section—the reader cannot assess whether the exact enforcement translates to better interior solution accuracy or faster convergence versus well-tuned soft constraints. The comparison to other exact-enforcement methods (e.g., approximate distance functions of Sukumar & Srivastava) correctly notes that those approaches struggle with Neumann/Robin conditions at non-smooth boundaries, which this method addresses through explicit compatibility constraint handling.

The mathematical formulation is presented in sufficient detail for reproduction by experts: the TFC constrained expressions (equations 15, 35, 63, 92), transfinite interpolation operators (P defined in equation 8), and Hermite polynomial definitions (equations 24, 49) are all explicit. Hyperparameters ($R_m$, $Q$, $M$, network architecture $[2, M, 1]$) are reported for each test case. However, no code repository or data is explicitly provided, and the implementation involves non-trivial derivative calculations through the Jacobian mapping (equations 102a-102b) that would benefit from released reference code. The ELM training avoids backpropagation in favor of linear/nonlinear least squares, making the training deterministic given fixed random seed for hidden layer initialization.